Traduction automatique et apprentissage d'une langue seconde

À partir de deux présentations qui appliquent l'analyse d'erreurs à des langues et des systèmes de traduction automatique différents, il s'agira de réfléchir aux usages possibles de la traduction automatique dans l'enseignement des langues.

A Pilot Study on Machine Translation Post-editing Training in Foreign Language Learning
Présentation de Hong Zhang, doctorante à l'université autonome de Barcelone.

The main purpose of this study is to show the usefulness of Machine Translation (MT) and Post-Editing (PE) training for Chinese students learning Spanish as a second language. Our hypothesis is that with specific MTPE training, foreign language students are able to detect and correct MT mistakes in the foreign language. We have developed Post-Editing (PE) training material based on the results collected in a previous study. The training material shows six typical error types of the raw output of Machine Translation (MT), and our Chinese students studying Spanish could not detect some of them in the experiment last year.  These error types are as follows: Accuracy; Word Order; Proper Noun (for named entity); Preposition; Omission; Formal Register. The training material also includes three levels of difficulty: 1) Identification of the MT error; 2) Error type classification and 3) Error resolution.
This year, another pilot study was performed in order to evaluate the effectiveness of the PE training materials. Sixteen participants were Chinese speakers with an intermediate knowledge of Spanish. Pretest - Posttest design was chosen. In the Pretest, all sixteen participants were asked to post-edit the Google Translate raw output from Chinese to Spanish. Then they were divided into two groups -- an experimental group and a control group. The experimental group performed 4 hours PE training and control group did not. In the posttest, both groups did the PE of the same raw output from the pretest. When comparing the results of the two groups, the experimental group could identify and correct more errors successfully than the control group. The experimental group performed well on Omission, Word Order, and Proper Nouns after the PE training. They were specifically more successful in edits and used less time than control group in the posttest.
From the results of this study, we can see that PE training can help Chinese students to identify and correct the errors of Omission, Word Order and Proper Nouns in the raw output of MT from Chinese to Spanish.  However, it seems that during the short PE course, the errors of Accuracy, Preposition and Formal Register were difficult for them. In future, we can focus on what the students struggle with in their Spanish studies and then expand the PE training materials.

Un outil pour apprendre de ses erreurs : ACCOLÉ. Les erreurs comme point commun entre la Traduction Automatique, les traducteurs et les apprenants.
Présentation d'Emmanuelle Espérança-Rodier, LIG-GETALP.

Nos recherches visent à établir un schéma typique des différents systèmes de Traduction Automatique (TA) qu'ils soient neuronaux ou probabilistes, afin d'orienter vers les bon outils de TA les personnes ayant besoin d'une traduction pour une tâche spécifique (traduction d'une page web, traduction d'un article scientifique etc…) et en fonction de leur compétence dans la langue cible et le domaine du texte à traduire.
Nous avons créé une plateforme en ligne, ACCOLÉ, qui permet non seulement d'annoter des erreurs de traduction en fonction de typologies prédéfinies comme celles de Vilar et al. (2006), orientées linguistique, celles des métriques multidimensionnelles de qualité (MQM) (Lommel, 2018), orientées outils, ou celle qui a été créée dans le cadre du projet MeLLANGE (Castagnoli et al., 2011), orientée traduction, mais aussi de collaborer entre annotateurs afin d'améliorer l'accord inter-annotateur et surtout de pouvoir effectuer différentes recherches sur les corpus annotés afin de définir des patrons d'erreurs typiques des systèmes de TA étudiés.